SCIENCE UNDER AUTOMATION
GeoNatureAgent v2
93 tasks, 18 categories, 16 tools; Claude Sonnet 4 capability 60.8% ± 0.8%.
What this result supports
Production-style API for three environmental indicators in Spain and Portugal; does not test field sampling or ecological causal inference. Reported ± is not relabeled as a confidence interval.
Evidence record: preprint. Checked: 2026-10-20.
Where this evidence is used
Field ecology
- Translate a supplied environmental query: GeoNatureAgent evaluates supplied queries and geospatial tools.
- Retrieve and analyse available data: Performance is partial across 93 tasks in two countries.
- Interpret ecological implications: API success does not establish causal or sampling validity.
- Design representative sampling: Geospatial API tasks do not test field sampling design.
- Collect data in variable environments: Physical fieldwork is not evaluated.
- Distinguish causal change and validate it: Independent causal ecological inference is not evaluated.